Reverb/arabic-eou-conversations

Arabic EOU Conversations Dataset

Multi-dialect Arabic conversation dataset for End-of-Utterance (EOU) detection.

Dataset Description

This dataset contains Arabic conversational examples labeled for end-of-utterance detection. It includes natural dialogue patterns across three major Arabic dialects.

Key Features:

• Total Examples: 11,660 (balanced 50/50 EOU/NOT_EOU)
• Dialects: Levantine (شامي), Egyptian (مصري), Gulf (خليجي)
• Format: Conversation pairs with EOU labels
• Split: 80% train, 10% validation, 10% test

Dataset Structure

Each example contains:

• text: The utterance text in Arabic
• label: Binary label (0 = NOT_EOU, 1 = EOU)
• context: Optional previous utterance for context

Example

{
  "text": "شو رأيك نروح نتغدا؟",
  "label": 1,
  "context": ""
}

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("Reverb/arabic-eou-conversations")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

# Example
print(train_data[0])

Data Collection

The dataset was curated from Arabic conversational data across three dialects:

• Levantine: Conversational Arabic from the Levant region
• Egyptian: Egyptian colloquial Arabic
• Gulf: Gulf Arabic dialects

Intended Use

This dataset is designed for:

• Training EOU detection models
• Turn-taking research in Arabic dialogue
• Voice assistant development
• Conversational AI systems
• Arabic NLP research

Limitations

• Focused on conversational/colloquial Arabic
• May not represent all Arabic dialects equally
• Designed for spoken language patterns
• Not suitable for formal/literary Arabic

Citation

@misc{arabic-eou-dataset,
  author = {Reverb},
  title = {Arabic EOU Conversations Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Reverb/arabic-eou-conversations}
}

License

MIT License

Related Models

This dataset was used to train: Reverb/smollm2-135m-arabic-eou